Pure and Applied Mathematics Quarterly Volume 13, Number 2, 215–246, 2017 Complete complexes and spectral sequences Mikhail Kapranov and Evangelos Routis To Yuri Ivanovich Manin Abstract: By analogy with the classical (Chasles-Schubert-Semple- Tyrell) spaces of complete quadrics and complete collineations, we introduce the variety of complete complexes. Its points can be seen as equivalence classes of spectral sequences of a certain type. We prove that the set of such equivalence classes has a structure of a smooth projective variety. We show that it provides a desin- gularization, with normal crossings boundary, of the Buchsbaum- Eisenbud variety of complexes, i.e., a compactification of the union of its maximal strata. 0 Introduction 215 Acknowledgements 218 1 Categories of complexes 218 2 The affine variety of complexes, its strata and normal cones221 3 The projective variety of complexes 225 4 The relative varieties of complexes and their normal cones 226 5 Charts in varieties of complexes 228 6 Complete complexes via blowups 232 7 Complete complexes and spectral sequences 240 References 245 0. Introduction A. Background and motivation. The spaces of complete collineations and complete quadrics form a beautiful and very important chapter of algebraic Received 29 September 2017. 215 216 Mikhail Kapranov and Evangelos Routis geometry, going back to the classical works of Chasles and Schubert in the 19th century, see [Se], [Ty], [L3], [L2], [Va], [Th], [DGMP], [Kau] and refer- ences therein. They provide explicit examples of wonderful compactifications (i.e., of smooth compactifications with normal crossings boundary). ∗ To recall the basic example, the group PGLn(C)=GLn(C)/(C · 1)has an obvious compactification by the projective space P(Matn(C)) but it is not wonderful since the complement, the determinantal variety, is highly singular. Now let V,W be two C-vector spaces of the same dimension n.Acomplete collineation from V to W is a sequence of the following data (assumed to be nonzero and considered each up to a non-zero scalar factor): (0) A linear operator A = A0 : V → W , possibly degenerate (not an iso- morphism). Note that Ker(A)andCoker(A) have the same dimension. (1) A linear operator A1 :Ker(A0) → Coker(A0), possibly degenerate. (2) A linear operator A2 :Ker(A1) → Coker(A1), possibly degenerate, and so on, until we obtain a non-degenerate linear operator. One of the main results of the classical theory says that the set of complete collineations has a natural structure of a smooth projective variety over C, containing PGLn(C) as an open subset (A0 non-degenerate) so that the com- plement is a divisor with normal crossings. We now want to look at this classical construction from a more modern perspective. We can view a linear operator A : V → W as a 2-term cochain complex. Then the sequence (Aν ) as above is nothing but a spectral sequence: • ν • a sequence of complexes (Eν ,D ) such that each Eν+1 is identified with the • • cohomology HDν (Eν ). This suggests a generalization of the construction of complete collineations involving more full-fledged (simply graded) spectral sequences. In this paper we develop such a generalization. The role of the group GLn(C)(oritspro- jectivization PGLn(C)) is played by appropriate strata in the Buchsbaum- Eisenbud variety of complexes C(V •) and its projectivization PC(V •). Here V • is a graded vector space and C(V •) consists of all ways of making V • into a cochain complex, see §2 and [Ke], [DS] for more background. The vari- eties C(V •) are known to share many important properties of determinantal varieties, in particular, they are spherical varieties: the action of the group GL(V •)= GL(V i) on the coordinate ring has simple spectrum, i.e., each irreducible representation enters at most once. B. Summary of results. Our results can be summarized as follows. For simplicity, consider the projective variety of complexes PC(V •). Let PC◦(V •) Complete complexes and spectral sequences 217 be the union of its maximal GL(V •)-orbits, a smooth open dense subvariety in PC(V •), see (6.7). AtthesametimeletPSS(V •) be the set of equivalence classes of spectral • ν ··· sequences (Eν ,D ), ν =0, ,N, of variable (finite) length N,see§7A, in which: • • • E0 = V . • Each Dν, ν =0, ··· ,N − 1, is not entirely zero and considered up to an overall scalar. • • The “abutment” EN does not admit any two consecutive nonzero spaces • (so the spectral sequence must degenerate at EN ). Then: (1) The set PSS(V •) admits the structure of a smooth projective variety PC(V •)overC. (2) PC(V •)containsPC◦(V •) as an open dense part, and the complement PSS(V •) − PC◦(V •) is a divisor with normal crossings. (3) One can obtain PC(V •) as the successive blowup of the closures of the natural strata in PC(V •). These results are obtained by combining Theorems 6.10 and 7.3. The realiza- tion of PC(V •) as an iterated blowup generalizes the approach of Vainsencher [Va] to complete collineations. In the main body of the paper we work over any algebraically closed field k of characteristic 0 and consider the varieties C(V •) as well. Also, more generally, for any graded locally free sheaf of finite rank V • over an arbitrary normal variety X over k,weintroducerelativever- • • sions of varieties of complexes CX (V )andPCX (V ) (cf. Section 4). When X is smooth we obtain the analogs of the results (1)-(3) above. C. Phenomena behind the results. The main phenomenon that makes our theory work, is the remarkable self-similarity of the variety of complexes.More precisely, C(V •) is subdivided into strata (loci of complexes with prescribed ranks of the differentials). The transverse slice to a stratum passing through apointD ∈ C(V •) (i.e., a differential in V •), is itself a variety of complexes • • but corresponding to the graded vector space HD(V ) of cohomology of D (cf. Propositions 2.9 and 4.4). This generalizes the familiar self-similarity of the determinantal varieties: the transverse slice to the stratum formed by matrices of fixed rank inside a determinantal variety, is itself a determinantal variety of smaller size. In particular, our analysis implies that our stratification is conical in the sense of MacPherson and Procesi [MP]. Further, the classical intuitive reason behind the appearance of complete collineations has a transparent homological meaning. To recall this reason, 218 Mikhail Kapranov and Evangelos Routis consider a 1-parameter family A(t) of linear operators V → W (depending, say, analytically on a complex number t near 0). If A(t) is nondegenerate for t = 0 but A0 = A(0) is degenerate, then the “next Taylor coefficient” of A(t) gives A1 :Ker(A0) → Coker(A0), the further Taylor coefficients give A2 and so on. This gives the limit limt→0 A(t) in the space of complete collineations in the classical theory. If we now have an analytic 1-parameter family D(t) of differentials in the same graded C-vector space V •, we can view the Taylor expansion of • D(t) as a single differential D in the graded C((t))-vector space V ⊗C C((t)). The fact that D(t) is analytic at 0 (so we are talking about Taylor, not • Laurent expansions), means that D preserves the t-adic filtration in V ⊗C C((t)). The associated spectral sequence of the filtered complex is essentially simply graded, and it represents the limit of (V •,D(t)), as t → 0, in our compactification. D. Future directions. We expect our varieties of complete complexes to have interesting enumerative invariants, generalizing the many remarkable properties of complete collineations. Historically, the first example of a “complete” variety of geometric objects was the Chasles-Schubert space Qn of complete quadrics, which gives a won- derful compactification of the variety Qn of smooth quadric hypersurfaces in n P ,see[DGMP]. From our point of view, Qn can be seen as a particular case of the variety of self-dual complexes. That is, we start with a graded (by Z or Z 1 • + 2 ) vector space V which is identified with its graded dual by a graded symmetric bilinear form and consider all ways of making V • into a self-dual complex. The corresponding analog of Qn is then formed by the variety of self-dual spectral sequences. We leave its study to a future work. Acknowledgements This work was supported by the World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan. 1. Categories of complexes Let k be an algebraically closed field and Λ be a finitely generated commuta- tive k-algebra. We denote by gModΛ the category of finitely generated graded • • i i graded Λ-modules V .Thatis,V = i∈Z V , with all V finitely generated, and V i =0for|i|0. For j ∈ Z the shifted graded module V •[j] is defined by (V •[j])i = V j+i. Complete complexes and spectral sequences 219 C We denote by omΛthe category of cochain complexes over Λ, i.e., of • i ∈ graded modules V = V gModΛ equipped with a differential D,a i i+1 collection of Λ-linear maps Di : V → V satisfying Di+1 ◦ Di =0.We • → • will consider D =(Di) as a morphism V V [1] in gModΛ. For a complex • • • (V ,D) we have the graded module HD(V ) of cohomology. We define the shifted complex (V •,D)[j] to have the underlying graded • module V •[j] as above and the differential DV [j] having components given by • V [j] − j V • (1.1) Di =( 1) Dj+i. • • One way to explain this formula is to represent V [j]=Λ[j] ⊗Λ V (here Λ[j] is the ring Λ put in degree (−j)).